Abstract: Abstract: To address the problem that coronary artery calcified plaques in computed tomography angiography (CTA) images are small, sparsely distributed, and are easily confused with high-density structures such as contrast-filled lumens, bones, and stents, which leads to missed detections and false positives, a coronary artery calcified plaque segmentation method based on anatomical priors and Multi?Scale Context Refinement Network (MSCR?Net) was proposed in this paper. This method adopts a two-stage cascaded segmentation framework. First, a coronary lumen probability map was generated by our previously developed nnU-Net v2-based lumen segmentation network, serving as the anatomical spatial prior for the second stage. Second, MSCR?Net was constructed as a dual-input network, fed with raw CTA volumes and the precomputed coronary prior map, where the prior constrained the search space for calcifications. In the shallow layers, a local multi?scale residual encoder was adopted to extract fine-grained textures and local context; in the bottleneck, a customized Atrous Spatial Pyramid Pooling (ASPP) combined with a Convolutional Block Attention Module (CBAM) was employed to aggregate multi-scale semantic information and suppress high-density background interference; and in the decoder, a progressive refinement strategy was used to gradually recover boundary details. Finally, a spatially weighted joint loss integrating Tversky loss and Focal loss was introduced, together with deep supervision and prior-guided post-processing, to alleviate severe foreground-background imbalance and improve the learning of hard samples. Experimental results on 139 high-resolution CTA scans demonstrate that the proposed method achieves a Dice Similarity Coefficient (DSC) of 75.99%, outperforming nnU-Net by 4.33 percentage points, and reduces the 95% Hausdorff Distance (HD95) to 16.55 mm, which is 6.32 mm lower than that of Residual U-Net, with only 15.73 M parameters, while maintaining relatively low computational cost and achieving superior segmentation accuracy and boundary localization performance. The proposed method can provide a reliable imaging reference for automated quantitative analysis of coronary calcification and facilitate computer-aided diagnosis and risk assessment of cardiovascular diseases (CVD).

Key words: Keywords: coronary artery calcified plaque segmentation, Computed Tomography Angiography (CTA), multi-scale context, attention mechanism, coronary anatomical prior

摘要： 摘 要: 针对冠状动脉计算机断层扫描血管造影（CTA）图像中钙化斑块体积小、分布稀疏，且易与造影剂充盈管腔、骨骼及支架等高密度结构混淆，导致漏检与误判的问题，本文提出一种基于解剖先验与多尺度上下文精化网络（MSCR-Net）的冠脉钙化斑块分割方法。该方法采用两阶段级联分割框架：首先，采用课题组前期提出的基于nnU-Net v2的冠脉管腔分割网络生成冠状动脉管腔概率图，作为解剖空间先验；其次，构建MSCR-Net双通道输入网络，以原始CTA图像和冠脉先验概率图为输入，利用先验图约束钙化搜索空间，在浅层通过局部多尺度残差编码器提取细粒度纹理和邻域上下文，在瓶颈层通过改进的空洞空间金字塔池化（ASPP）与卷积块注意力模块（CBAM）聚合多尺度语义信息并抑制高密度背景干扰，在解码端采用渐进式精化策略逐步恢复边界细节；最后，引入空间加权的Tversky与Focal联合损失，并结合深度监督和基于先验的后处理策略，以缓解前景-背景严重失衡并提升难分样本的学习能力。基于139例高分辨率CTA数据的实验结果表明，本文方法的Dice相似系数（DSC）达到75.99%，较nnU-Net提升了4.33个百分点；95% Hausdorff距离（HD95）为16.55 mm，较Residual U-Net降低了6.32 mm，且参数量仅为15.73M，在保持较低计算开销的同时取得了最优的分割精度与边界定位能力。本文方法可为冠状动脉钙化斑块的自动定量分析提供可靠的影像学参考，助力心血管疾病（CVD）的计算机辅助诊断与风险评估。

关键词: 关键词: 冠状动脉钙化斑块分割, 计算机断层扫描血管造影, 多尺度上下文, 注意力机制, 解剖先验